Ambit

What ambit does

The ambit MCP server turns scattered agent configuration into a capability graph. It reads configurations from runtimes such as Claude Code, Cursor, OpenCode, and other supported environments, then combines the discovered MCP servers, models, skills, providers, declared credentials, commands, and machines with Ambit's curated capability tree.

The resulting graph separates several states that configuration files alone do not establish. A capability can be structurally reached because its provider and prerequisites are present, verified because its declared check passed recently, or blocked because a required prerequisite is missing. Configured items that fail verification are excluded from plans and permissions until they pass again.

The graph also identifies the frontier: capabilities that are one setup step away. It can account for combinations assembled from separate tools, such as a vector store and local embeddings contributing to semantic retrieval. This makes the ambit MCP server useful for both small agent setups seeking their next missing component and larger environments where shared dependencies or broken configurations need review.

How it works

Ambit builds a local model from the configuration sources it knows how to read. Nodes represent configured resources and curated capabilities; edges represent requirements and provided capabilities. Dependency tracing includes transitive relationships, so several providers sharing one credential can be identified as depending on a common failure point.

The MCP interface lets agents ask what they can do before attempting a task, what prerequisite is missing when they reach a limit, and what change could expand their reach. Proposed changes wait for approval rather than being applied automatically. Ambit also considers declared verification checks and can report whether a capability is installed, operational, or authorized.

From the command line, ambit status reports verified capabilities, ambit goal presents next steps, and ambit verify runs declared checks. The ambit web command serves the local capability map, including the visual view of reached, next-step, and blocked nodes.

Setup and configuration

Run the CLI with Node.js 22.18 or newer using npx ambit-cli, or install it through Homebrew with brew install zz-plant/tap/ambit && ambit. A global npm installation is also supported with npm install -g ambit-cli. After installation, ambit web serves the map locally.

The project can alternatively be checked out and started with its bootstrap script: git clone https://github.com/zz-plant/ambit.git, followed by ./bootstrap.sh web from the repository directory. A hosted browser demo is available for sample data, and the browser can accept configuration from supported runtimes without uploading it. The MCP server is part of the local CLI installation and can be registered with an agent using that agent's MCP configuration.

Tools and capabilities

The ambit MCP server supports these grounded workflows:

• Inspect the combined capability map for an agent setup.
• Ask which capabilities are reached, blocked, or one step away.
• Find missing prerequisites and estimate the setup time for next steps.
• Check configured capabilities with declared verification commands.
• Trace the downstream impact of a failed dependency.
• Rank possible setup work by blocked capabilities, unlocked reach, and recorded human attention where available.
• Draft configuration changes for approval rather than silently applying them.

The command-line surface includes ambit status, ambit goal, ambit verify, and ambit web. The visual map shows dependencies and capability states, while the MCP surface makes the same model available to agents.

Limitations and notes

Ambit only understands the agent runtimes, configuration formats, capability definitions, and verification checks it knows about. A capability marked as reached is not necessarily proven operational until its check passes. Conversely, unverified capabilities remain available in the model, while failed checks remove them from plans and permissions.

The graph describes declared configuration and dependencies; it does not by itself establish that an external service will remain available or that an authorization grants more permission than the configuration indicates. Proposed changes require human approval. The project requires Node.js 22.18 or newer when run through its Node-based CLI.